A society of specialist agents that improve a scored scientific model in parallel — edit, evaluate, keep or revert — while you steer and decide what’s real.
Demo target: a virtual cell (gene-perturbation response, held-out Pearson r). The claim: a parallel society beats a lone agent, and the run stays legible enough to trust.
AutoCell runs several agents at once, all improving the same model.
Each agent:
- picks a model from a shared archive,
- changes one part of it,
- trains and scores the result,
- writes it back to the archive.
The agents never message each other. The archive is their shared memory, so a later agent builds on what an earlier one found.
Each role runs on a different Qwen model, matched to the work.
Agents can edit only the model file: its features, architecture, loss, and optimization. The data and the scoring metric stay out of reach, so the only way to score higher is to build a better model.
When a result looks too strong, AutoCell holds a review. One agent defends the change, another re-tests it on different data, and a third decides.
A web console lets you follow the experiments, inspect a change, watch the debate, and keep or drop the result. The review can also run aloud, over Qwen's real-time voice model.
Equal-budget comparison in results/step2/:
| Society | Solo | Gap | |
|---|---|---|---|
| Mock (mechanism) | 0.280 | 0.238 | +0.042 |
| Qwen primary | 0.2354 | 0.2348 | +0.0005 |
| Qwen robust (3 seeds) | 0.2226 | 0.2193 | +0.0034 |
The mock shows the search shape working when diversity is rewarded. The real Qwen edge is thin on primary and clearer on the robust k-split — reported honestly.
# key — never commit it
cp .env.example .env # set DASHSCOPE_API_KEY
# API (serves studies, meetings, live-voice proxy)
pip install -r requirements.txt
uvicorn server.app:app --reload --port 8000
# console
cd web && npm install && npm run dev
# open http://localhost:5173 — mock UI works with VITE_API unset
# for the real archive: VITE_API=http://127.0.0.1:8000Society vs solo (torch-free mock, no dataset):
python -m bench.run --mock --mode society --budget 24
python -m bench.run --mock --mode sequential --budget 24Deploy the API to Alibaba ECS (key injected at runtime only):
export ECS_HOST=<public-ip>
bash deploy/deploy.shprepare.py frozen data + metric
train.py the one editable model file
core/ archive, sampler, proposer, runner
bench/ society-vs-solo measurement harness
society/ meeting, cast, steering
llm/ Qwen client + realtime session
server/ FastAPI
web/ React console
results/step2/ the proof numbers
diagram/ architecture figures
deploy/ Dockerfile + ECS script
Python · FastAPI · SQLite · React/Vite · PyTorch (train.py) · Qwen Cloud (DashScope, OpenAI-compatible)